A lightweight P2P-based cache system for model distributions on Kubernetes.

Name Story: the inspiration of the name Manta is coming from Dota2, called Manta Style, which will create 2 images of your hero just like peers in the P2P network.

We're reframing the Manta to make it a general distributed cache system with POSIX promise, the current capacities are still available with the latest v0.0.4 release. Let's see what will happen.

Architecture

Note: llmaz is just one kind of integrations, Manta can be deployed and used independently.

Features Overview

• Model Hub Support: Models could be downloaded directly from model hubs (Huggingface etc.) or object storages, no other effort.
• Model Preheat: Models could be preloaded to clusters, or specified nodes to accelerate the model serving.
• Model Cache: Models will be cached as chunks after downloading for faster model loading.
• Model Lifecycle Management: Model lifecycle is managed automatically with different strategies, like Retain or Delete.
• Plugin Framework: Filter and Score plugins could be extended to pick up the best candidates.
• Memory Management(WIP): Manage the reserved memories for caching, together with LRU algorithm for GC.

You Should Know Before

• Manta is not an all-in-one solution for model management, instead, it offers a lightweight solution to utilize the idle bandwidth and cost-effective disk, helping you save money.
• It requires no additional components like databases or storage systems, simplifying setup and reducing effort.
• All the models will be stored under the host path of /mnt/models/
• After all, it's just a cache system.

Quick Start

Installation

Read the Installation for guidance.

Preheat Model

A sample to preload the Qwen/Qwen2.5-0.5B-Instruct model. Once preheated, no longer to fetch the models from cold start, but from the cache instead.

apiVersion: manta.io/v1alpha1
kind: Torrent
metadata:
  name: torrent-sample
spec:
  hub:
    name: Huggingface
    repoID: Qwen/Qwen2.5-0.5B-Instruct

If you want to preload the model to specified nodes, use the NodeSelector:

apiVersion: manta.io/v1alpha1
kind: Torrent
metadata:
  name: torrent-sample
spec:
  hub:
    name: Huggingface
    repoID: Qwen/Qwen2.5-0.5B-Instruct
  nodeSelector:
    foo: bar

Use Model

Once you have a Torrent, you can access the model simply from host path of `/mnt/models/. What you need to do is just set the Pod label like:

metadata:
  labels:
    manta.io/torrent-name: "torrent-sample"

Note: you can make the Torrent Standby by setting the preheat to false (true by default), then preheating will process in runtime, which obviously wll slow down the model loading.

apiVersion: manta.io/v1alpha1
kind: Torrent
metadata:
  name: torrent-sample
spec:
  preheat: false

Delete Model

If you want to remove the model weights once Torrent is deleted, set the ReclaimPolicy=Delete, default to Retain:

apiVersion: manta.io/v1alpha1
kind: Torrent
metadata:
  name: torrent-sample
spec:
  hub:
    name: Huggingface
    repoID: Qwen/Qwen2.5-0.5B-Instruct
  reclaimPolicy: Delete

More details refer to the APIs.

Roadmap

In the long term, we hope to make Manta an unified cache system within MLOps.

• Preloading datasets from model hubs
• RDMA support for faster model loading
• More integrations with MLOps system, including training and serving

Community

Join us for more discussions:

• Slack Channel: #manta

Contributions

All kinds of contributions are welcomed ! Please following CONTRIBUTING.md.